Trunk miscompiles libjava (resend)

["Boehm, Hans" <hans_boehm at hp dot com>]

I attempted to send this to both java and gcc lists yesterday, but there seems to have been a mailer glitch ...

-----Original Message-----
From: Boehm, Hans
To: 'gcc@gcc.gnu.org'; 'java@gcc.gnu.org'
Sent: 5/17/02 7:30 PM

On Linux/IA64, and probably other platforms, the trunk gcc miscompiles
the isWhiteSpace() method in libjava/java/lang/Character.  Apparently
the bit-wise & in a return is the problem.  This is a regression from
3.1.  It fails with the current CVS and failed with the one from a few
days ago.

Reduced test case (only isWhitespace is miscompiled, the rest is
scaffolding):

class mychar {
  public static boolean isWhitespace(char ch)
  {
    int attr = readChar(ch);
    return (((1 << attr) & 0x7000) != 0);
  }


  private static char readChar(char ch) {
    if (ch < 128) return (char) 9;
    return 9999;
  }

  public static void main(String[] argv) {
    if (isWhitespace('4'))
      System.out.println("FAIL 1");
    if (isWhitespace('a'))
      System.out.println("FAIL 2");
  }
}

Offending RTL instruction (partial diff of trunk (-) against 3.1 output
(+)):

 (insn 57 56 59 (set (reg:DI 359)
-        (zero_extend:DI (subreg:QI (reg:SI 355) 0))) -1 (nil)
+        (and:DI (subreg:DI (reg:SI 355) 0)
+            (const_int 1 [0x1]))) -1 (nil)
     (nil))
 
In both 3.1 and the trunk, the code in the return is compiled as a
variable right shift of 0x7000.  In the 3.1 version, this is anded
against 0x1.  In the trunk version the and is dropped.

I didn't yet track this down much further, but it fails at -O0, and the
initial rtl is already wrong.  It works correctly if I compile to a
class file and then interpret with gij.  Any ideas?

Hans